New research presented on Thursday explored whether deep learning models using artificial intelligence (AI) can determine a woman's risk for mammography screening-detected breast cancer or interval breast cancer, with and without the presence of clinical risk factors.
"There is an opportunity to target risk reduction or screening strategies with more precision if we understand a woman's likelihood of developing either screening-detected or interval breast cancer," said Lambert Leong, a graduate student at the University of Hawaii Cancer Center and University of Hawaii Department of Molecular Bioscience and Bioengineering.
According to Leong, the deep learning model he and his fellow investigators used in the study outperformed clinical risk factors in determining the risk of screening-detected cancer.
In the case-control study, the researchers prospectively collected 25,096 negative digital screening mammograms. The mammograms were performed from 2006 to 2014, a median of 2.8 years before any cancer. Of these, 4,409 women were cancer-free, 351 had interval invasive breast cancer and 1,609 had screening-detected breast cancer.
The investigators trained and validated a deep learning model on the negative mammograms before cancer to classify the women as having developed screening-detected or interval invasive cancer, or not having developed cancer.
"AI was our method of choice because we had a fairly high number of cases to train the model and AI doesn't require an a priori hypothesis on what imaging feature or features is/are important," Leong said. "Showing how our AI model worked with and without clinical risk factors allowed us to evaluate the unique contribution of the AI model."
The researchers benchmarked the deep learning models against models using breast density and other clinical risk factors. They also adjusted all models for clinical risk factors including race, history of biopsy, first degree family history of breast cancer, age (continuous) and BMI (continuous, inverse BMI).
Predicting Cancer Risk
After adjusting for age, BMI, history of breast cancer, history of biopsy and race, deep learning models were able to better predict screening-detected risk when compared to clinical risk factors alone, Leong said.
And yet, the clinical risk factor model performed as well as deep learning for identifying interval cancer risk.
"Better deep learning performance, in the screen-detected case, indicates the presence of imaging features of risk not captured with breast density alone," Leong continued. "In the interval cancer instances, however, breast density appeared to be the main driver of risk predictions. We interpret this as indicating that masking due to opacity is the main driver of masking of cancers associated with interval cancer risk. There are, of course, alternative explanations that also need to be explored."
The results of this and similar studies, Leong noted, could lead to women at high risk of screening-detected cancers having more frequent mammography screenings and women at high risk of interval breast cancer opting for alternative screening strategies, thereby helping to find cancer earlier and improve patient outcomes.
Access the presentation, "Deep Learning Predicts Interval And Screendetected Cancer From Negative Screening Mammograms: A Case-case-control Study In 6,369 Women," (SSBR10-5) on demand at Meeting.RSNA.org.